AI and Health Equity: What Language Access Reveals About Safe AI

The most evident and oldest issue analyzing if healthcare AI is equitable is if the patient and the clinician can communicate effectively. In a Signal Room conversation, Carol Velandia — founder of Equal Access Language Services and a trainer to healthcare personnel about language and communication — explained where the risk lies in this gap. She argued that communication is the most valid and vital diagnostic tool. When communication is somehow compromised, for example, by a language barrier, everything in the continuum of care — from diagnosis to discharge — becomes suboptimal and unnecessarily costly.

This issue does not belong to a particular category. Velandia cited research finding that people with limited English proficiency are significantly more likely than English speakers to suffer serious temporary harm, in addition to the tens of millions of people in the U.S. who must deal with the healthcare system using a language that is not English. At Hutchins Data Strategy Consultants, language access is a useful framework for researching a broader issue — whether an AI service is truly accessible to the entire population it intends to serve.

A Civil Right, Not a Courtesy

Velandia's argument serves to counter the tendency of most framing to place language access as a final, outward-facing defensible courtesy for an organization. She uses the analogy of access to a building via its entrance. From an abstract perspective, a staircase is unobtrusive, and therefore unproblematic. However, if a person uses a wheelchair, a staircase is a building access barrier. She claimed that, for the millions of people who experience limited proficiency in English, the barrier of English-only care is a steep, tiring, and at times, a deadly burden. Unlike a missing ramp, the discrimination that this represents is not readily viewed as discrimination.

Her argument is law-based as well as principle-based. Through Title VI and Section 1557 of the Affordable Care Act, which carry provisions for staff training requirements, language access is a civil-rights law protection obligation. For Velandia's audiences (organizational and health care leaders), this builds access as part of a system's design rather than as a bedside improvisation. The way most organizations handle language access today — solving barriers in any available manner, with untrained interpreters or even children — is a system's failing the law's protection exists to prevent. The data she provided supports the claim: the presence of a professional interpreter, as opposed to an ad hoc one, leads to far fewer mistakes.

Where AI Helps — and Where It Breaks

Velandia does see a role for AI in translation. AI can help speed up the process and help translate text into more languages. That is a positive thing. But AI will never be able to verify the accuracy of translations in a clinical setting. AI will always make translation tasks faster, and not necessarily more accurate. This is where humans become an important factor.

To illustrate her point, she gave an example of a breakdown in the translation process. Translation of a clinical phrase that is constructed from the acronym "SAFETY" was translated literally and, as a result, made no sense in the target language. The problem was not an inadequate algorithm. The solution was a human who adjusted the translation to retain its meaning. The absence of a translation can be better than an inaccurate translation. There is a risk with a falsely accurate translation that there is a message that has been expressed, even though communication has not happened. This is a problem with medical text, as there are a lot of acronyms used which can carry a lot of meaning.

There is a risk with errors in translations compared to errors in procedures or misadministration of medication. Errors in translation can be even worse than errors in procedures or misadministration of medication, especially since the errors in translation are often not noticed, unlike errors in medication.

You Cannot Outsource Ethics

Regarding accountability, what can be drawn from Velandia's remarks reveals much beyond access to language. It is not possible to outsource ethics, she argues, to AI. An AI will not safeguard the dignity of a patient; it is the human being involved in the interaction who does so. If you were to isolate your reasoning to determine if the use of technology to facilitate communication is justified, consider her test. If you were situated in an environment where no one is able to communicate in a language that you understand, would you feel comfortable receiving a diagnosis, prognosis, and treatment entirely through a tablet, with no human in the conversation? The aim of her question is to illustrate the discomfort associated with reliance on technology. The speed with which technology can communicate, and the breadth of its reach, cannot substitute for the culturally grounded moral responsibility and human judgement which are essential to ensure the safety and the dignity of the patient in the care that is provided.

Her model relies on AI where the human element remains in the loop — in a partnership with professional interpreters and human post-editing — designed to communicate swiftly and broadly while remaining accurate and culturally respectful. It is the same human-in-the-loop principle that governs safe clinical AI, applied to this high-stakes act of making sure a patient understands what is happening to them.

The Broader Equity Lesson

One of the patterns present in the majority of the use cases for healthcare AI is language access. If you create technology that is trained on a small subset of a population, it will continue that narrowness in the care it provides and most likely exacerbate the inequities it was designed to address. Following Velandia's standard, if you would want a professional interpreter and a human clinician to be your standard, then another patient should not receive a lower one. AI that disregards the language of a population, as well as its history and context, is not only underperforming for the specific population, but is also deepening the inequities that exist.

How Hutchins Approaches AI and Health Equity

Our work treats equity as a design requirement, not a value statement. We help organizations build language and communication access into systems from the start, scope where AI genuinely extends reach and where a human must remain in the loop, and hold every deployment to the test of whether it serves the whole population rather than the convenient majority. This is the equity dimension of responsible AI and of clinical design that works in real conditions.

These conversations are explored throughout The Signal Room podcast, where advocates and clinicians describe what it actually takes for care to reach everyone.

Authoritative sources

• U.S. Department of Health & Human Services — Office for Civil Rights
• WHO — Ethics & governance of AI for health
• Agency for Healthcare Research and Quality
• NIST AI Risk Management Framework